v0.5.3

[Talmaj]

Summary of Changes

BatchNorm Fixes & Improvements

• Fixed critical bias bug: Corrected line 59 in batchnorm.py where bias was incorrectly set to scale instead of B
• Fixed inference mode for batch_size > 1: Explicitly set BatchNorm to eval mode to use running statistics (ONNX inference behavior) Inconsistency between pytorch and onnxruntime #44
• Removed experimental flag: BatchNorm now works correctly with batch_size > 1 by default Why batch_norm will give false outputs when batch_size>1 #35
• Added comprehensive tests: Validated against onnxruntime with various batch sizes (1, 2, 4, 8), channels, spatial dimensions, epsilon values, and momentum values

New Operators

ReduceSumSquare

• Implemented ReduceSumSquare operator: Computes sum(x^2) along specified axes
• Supports both opset versions: Handles axes as attribute (opset < 13) and as optional input (opset >= 13)
• Comprehensive test coverage: 16+ parametrized test cases validating against onnxruntime with different input shapes, axes, and keepdims settings

LogSoftmax

• Added LogSoftmax operator: Supports axis attribute with proper default handling (dim=-1) Conversion request for op_type=LogSoftmax. #49 PR to add LogSoftmax, and also has some code cleanup #30
• Added comprehensive tests: Each operator validated against onnxruntime with various input shapes, axes, and mathematical property verification